Spoken Language Understanding for Task-oriented Dialogue Systems with Augmented Memory Networks

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 797–806 June 6–11, 2021. ©2021 Association for Computational Linguistics 797 Spoken Language Understanding for Task-oriented Dialogue Systems with Augmented Memory Networks Jie Wu University of California Irvine wuj8@uci.edu Ian G. Harris University of California Irvine harris@ics.uci.edu Hongzhi Zhao Beijing Jiaotong University hzzhao@bjtu.edu.cn Abstract Spoken language understanding, usually in- cluding intent detection and slot filling, is a core component to build a spoken dialog sys- tem. Recent research shows promising results by jointly learning of those two tasks based on the fact that slot filling and intent detection are sharing semantic knowledge. Furthermore, attention mechanism boosts joint learning to achieve state-of-the-art results. However, cur- rent joint learning models ignore the following important facts: 1. Long-term slot context is not traced effectively, which is crucial for fu- ture slot filling. 2. Slot tagging and intent detection could be mutually rewarding, but bi- directional interaction between slot filling and intent detection remains seldom explored. In this paper, we propose a novel approach to model long-term slot context and to fully uti- lize the semantic correlation between slots and intents. We adopt a key-value memory net- work to model slot context dynamically and to track more important slot tags decoded be- fore, which are then fed into our decoder for slot tagging. Furthermore, gated memory in- formation is utilized to perform intent detec- tion, mutually improving both tasks through global optimization. Experiments on bench- mark ATIS and Snips datasets show that our model achieves state-of-the-art performance and outperforms other methods, especially for the slot filling task. 1 Introduction Task-oriented dialogue systems have attracted sig- nificant attention, which have been greatly ad- vanced by deep learning techniques. Traditionally, these dialog systems have been built as a pipeline, with modules including spoken language under- standing (SLU), dialog state tracking, action se- lection and language generation. Among these problems, SLU, including intention detection and slot filling (Tur and Mori, 2011), is a key yet chal- lenging problem to parse users’ utterances into se- Sentence Flights from Irvine to Seattle Intent Flight Slots O O B-fromloc O B-toloc Table 1: An example utterance annotated with its intent and semantic slots (IOB format). mantic frames in order to capture a conversation’s core meaning. Traditionally, intention detection is treated as a classification problem, whereas slot filling is usually defined as sequence labeling prob- lem, where In-Out-Begin (IOB) format is applied for representing slot tags as illustrated in Table 1. Given an utterance, SLU determines users’ inten- tion and maps it into predefined semantic slots. The input is a sequence of words, and the output is a sequence of predefined slot IDs. A specific intent is assigned for the whole sentence. In the traditional pipeline approach, intent de- tection and slot filling are implemented separately. However, separate modeling of those two tasks is insufficient to take full advantage of all su- pervised signals, as they share semantic knowl- edge. For example, …

Content truncated to 500 words